Advanced Analysis Methods for 3G Cellular ... - Semantic Scholar
Advanced Analysis Methods for 3G Cellular Networks Jaana Laiho, Kimmo Raivio, Pasi Lehtim¨aki, Kimmo H¨at¨onen, Olli Simula Abstract— The operation and maintenance of the 3G mobile networks will be challenging. These networks will be strongly service driven, and this approach differs significantly from the traditional speech dominated 2G approach. Compared to 2G, in 3G the mobile cells interact and interfere with each other more, they have hundreds of adjustable parameters, and they monitor and record data related to several hundreds of different variables in each cell. This paper shows that a neural network algorithm called the Self-Organizing Map (SOM) together with a conventional clustering method like the k-means can effectively be used to simplify and focus network analysis. It is shown that these algorithms help in visualizing and grouping similarly behaving cells. Thus, it is easier for a human expert to discern different states of the network. This makes it possible to perform faster and more efficient trouble shooting and optimization of the parameters of the cells. The presented methods are applicable for different radio access network technologies. Keywords— Network management, 3G cellular system, WCDMA, radio access network, artificial neural network, Self-Organizing Map, k-means clustering, data mining.
I. Introduction The mobile communication industry is currently shifting its focus from second generation networks (2G) towards the third generation networks (3G). The shift is not only related to the evolution of the access technology, but also to the vision of the development of service provisioning and service demands, customer expectations and customer differentiation. While current wireless networks still evolve and service providers bring new internet packet data services into the markets, an increasing number of operators and other wireless communication professionals are becoming familiar with the wideband code division multiple access (WCDMA) technology and prepare themselves for 3G services and networks. There will be a number of new challenges when shifting from the current 2G to the new 3G networks, many of them related to the design and the operation of true multi-service radio networks. An essential part of the new challenges is related to the provisioning, monitoring and optimization of the services. The number of network counters and measurements shall increase owing to the fact that instead of monitoring GSM voice only, one must concentrate on monitoring multi system and multiservice environment. The information space of future netJ. Laiho and K. H¨ at¨ onen are with Nokia Group, Finland K. Raivio, P. Lehtim¨ aki and O. Simula are with the Neural Networks Research Centre, Helsinki University of Technology, Espoo, Finland The study has been financed by Nokia Networks, Nokia Mobile Phones and National Technology Agency of Finland (TEKES) which is gratefully acknowledged.
works is manifold compared to that of today’s networks. This fact is the main driver pushing development towards advanced network analysis solutions, like the one presented in this paper. Such functionality is located in the network and service management layers of the Telecom Operations Map (TOM) framework [1]. Operating a cellular network is an iterative quality cycle process combining the network and service configuration and the related performance measurements. In this cycle, the overall end-to-end quality target is defined and the quality criteria and thresholds for key performance indicators (KPI) for each service type are determined. Network performance data is gathered from Network Management Systems (NMS), drive tests, protocol analyzers and/or customer complaints. The actual measured service performance is analyzed and the results are compared with the set targets. In case of conflict, corrective actions to the network configuration are carried out. Effective analysis methods are prerequisites for dynamic and successful operations. This paper concentrates on the analysis and visualization part of the quality cycle. Datadriven algorithms and data mining methods provide efficient tools for exploratory data analysis. Algorithms based on Artificial Neural Networks (ANNs) have proved to be especially suitable in highly complex and data intensive applications. The Self-Organizing Map [2] is one of the most popular neural algorithms due to its efficient visualization properties. The motivation for the introduction of neural analysis on the network performance data is to provide effective means to handle multiple KPIs simultaneously. Furthermore, effective analysis methods reduce operators’ trouble shooting efforts, speed up the cycle, and thus, the network utilization rate increase. Furthermore, the quality as experienced by the end user (QoE) becomes increasingly important owing to the fact that over provisioning quality is inefficient and expensive. On the other hand, customers suffering from low quality tend to change to the competitor’s network. Determining the QoE is about collection and combination of information from different domains (access networks, interfaces and core network etc.). In this paper, an application of the SOM in analyzing telecommunications networks is presented. Mobile cells of a WCDMA network are classified according to their performance [3], [4]. SOM applications in a cellular environment are new and thus the results are verified using analytical results and expert knowledge. The Self-Organizing Map is introduced in Section II. Radio access network analysis methods based on the SOM and the results of the analysis are presented in Section III. In Section IV, the same
network is analyzed using conventional methods. Methods and results of the novel SOM based analysis are evaluated in Section V. Finally, the usability of the new method is discussed in Section VI.
data samples and prototype vectors in the input space. The update step can be performed by applying
II. The Self-Organizing Map
where α(t) is the learning rate and hc (t, i) is the neighborhood function of the algorithm. The last term in the square brackets is proportional to the gradient of the squared Euclidean distance d(x, mi ) = ||x − mc ||2 . The learning rate α(t) ∈ [0, 1] is usually a monotonically decreasing function of time. A good candidate is α(t) = α0 (1 − t/T ), where α0 is the initial value for the learning rate and T is the total number of training iterations. A very frequently used form for the neighborhood function hc (t, i) is the Gaussian one centered on the winner map unit c: ||ri − rc ||2 (3) hc (t, i) = exp − 2σ(t)2
mi (t + 1) = mi (t) + α(t)hc (t, i)[x(t) − mi (t)]
A. SOM in data mining Data mining and exploration is an emerging area of research in artificial intelligence and information management. The objective of data mining is to extract relevant information from large amounts of data. Data mining and analysis tasks typically include clustering, classification, and regression of data, determining parameter dependencies, and finding various anomalies from the data. Artificial Neural Networks provide efficient tools in data mining and they have been successfully used in various data engineering applications. The Self-Organizing Map (SOM) is a widely used neural network algorithm [2]. It maps a high-dimensional data manifold onto a low-dimensional, usually two-dimensional, grid or display. The SOM has several beneficial features which make it a useful tool in data mining and exploration. The SOM follows the probability density function of the data and is, thus, an efficient clustering and quantization algorithm. However, the most important feature of the SOM in data mining is the visualization property. The topology preserving property of the SOM mapping results in a display inherently visualizing the clusters in the data. SOM based methods have been applied in the analysis of process data, for example, in steel and forest industry [5], [6], [7], [8].
(2)
where rc depicts the coordinates of the winner unit c, and ri denotes the coordinates of an arbitrary unit i on the discrete output lattice of the map and σ(t) is the width of the neighborhood. It is necessary that hc (t, i) → 0 when t → ∞ for the algorithm to converge. During learning, the learning rate and the width of the neighborhood function are decreased, typically in a linear fashion. The map converges to a stationary distribution, which approximates the density of the data. One step of the training algorithm of the SOM is illustrated in Fig. 1. The size of the SOM is 16 units, which have been arranged into a two-dimensional grid of 4 by 4 units. A data sample is marked with a cross; the black circles are the values of the prototype vectors before, and the gray circles after updating them towards the data sample. This kind of an update step is repeated iteratively during the training process. There exist two freely available software packages that include implementation of the SOM: SOM-PAK [9] and SOM Toolbox for Matlab [10].
B. SOM algorithm The basic SOM consists of a regular grid of map units or neurons. They are connected to the neighboring units using, for instance, a rectangular or hexagonal neighborhood. Each map unit, denoted here by i, is represented by a prototype vector mi . The dimension of prototype vectors is equal to the dimension of the input data. First, the prototype vectors are initialized, for instance, to random values. Then, during the training the values of the prototype vectors are adapted to follow the properties of the input data. Training of the SOM is divided into two alternating steps, typically thousands of times each. First, one data vector x from the training data set is randomly selected and the corresponding best-matching (winner) unit (BMU) c is determined. The prototype vector of the BMU, denoted by mc , is the one that is nearest to the data sample. In other words, it minimizes the Euclidean distance between x and mc : c = arg min ||x − mi || (1)
Fig. 1. An illustration of the SOM training.
i
C. SOM in Network Analysis
In the next step, the prototype vectors of the winner and its neighbors are moved towards the data vector. It should be noted that the neighborhood is defined in terms of the lattice structure, not according to the distances between
A behavior pattern of a cell at a certain instant is a set of indicator values that have been recorded at that instant. In network analysis, the SOM can be used to find 2
common method to do the balancing is to normalize the variance of each variable equal to one. After the normalization, the distribution of the data might be skewed if there are outliers in the data. If the average normal behavior is studied, the usual solution is to remove outliers or to replace them with estimated normal or correct values. If outliers carry interesting information, for example, as is the case in our study, where they can be signs of network problems that are searched for, it is possible to keep outliers but prevent their large values from dominating the analysis results. This can be done by using some sort of conversion function like tanh(x) (or log(x)) before the variance normalization. Such a function can decrease the effect of outliers and emphasize proper parts of the distribution. For example, the tanh function emphasizes small values at the expense of large values.
and show similarities between behavior patterns of cells. These indicators can include, for example, any subset of those indicators that are used in the traditional analysis of a network. Indicators that have been used in this paper, are shown in Table II. A set of n indicator values form an n-dimensional pattern vector. During the training phase a set of these vectors is used to train a SOM. During the training, a SOM approximates the distribution of pattern vectors in the n-dimensional space so that everywhere in the space, where there are vectors, there are nodes of a SOM map as well. A prototype vector can be a BMU for several data samples that are almost similar. The SOM gives a topology preserving mapping of the input. SOM nodes that are direct neighbors in a SOM grid pull themselves towards each other. So when the training has been done, most of the SOM nodes neighboring each other in the grid are placed next to each other also in the input space. Therefore, cell behavior patterns that have been mapped to neighboring SOM nodes usually resemble each other. When the SOM is visualized, similarly behaving cells can be spotted close to each other.
C. Clustering analysis In general, clustering is the grouping of similar samples together. In this work, clustering is also used to find groups of similarly behaving cells. The features used in mobile cell clustering and classification are computed with the method shown in Fig. 2. The data vectors of all the cells are clustered using a combination of the Self-Organizing Map and the k-means algorithm. At first, a SOM with M map units is trained using the data vectors. Next, the set of M codebook vectors of the SOM are clustered into several different numbers of clusters using the k-means algorithm [11]. The clustering process can be repeated several times for different values of k, for √ example, 100 clusterings for each value of k, 2 ≤ k ≤ M . The k-means clustering has to be run several times for each k because the algorithm gives different results depending on the initialization. The best clustering for each k minimizes the sum of squared errors. Several values of k have to be tested because the correct number of clusters is not known. The optimal value of k is defined using some clustering validity index, like the Davies-Bouldin index [12]. Such an index is able to tell us the best number of clusters for the current data set. In this algorithm, the SOM is used to quantize the data and to visualize the cluster structure in the data because it is much faster to find the clusters of SOM codebook vectors than the clusters of original data directly [13].
III. Network analysis using SOM In this section, a neural method for classification of mobile cells is presented. The method consists of the following phases: target selection, data preprocessing, clustering analysis and result interpretation. A. Target selection The first step in the process is the target definition. This includes the selection of the geographical area, network objects (base stations, radio network controllers, routers, etc.) and visualization task specification. The selection of network objects and the visualization task have a strong impact on the selection of the measurements and KPIs to be analyzed. Naturally each object in the network has its own specific measurements. The visualization task can be more generic or problem oriented. General performance analysis requires a different set of measurements than a specific trouble shooting case. B. Preprocessing Neural network methods are multivariate methods that study the combination of variables, i.e., their joint distribution. Before they can be applied to the data, the data has to be prepared for the analysis in a preprocessing phase. The main objective of the preprocessing phase is to ensure that the analysis methods are able to extract correct and needed information from the data. Preprocessing can filter out noise, handle the problem of missing values and balance different variables and their value ranges. The action required stems from the current information need. For example, in network analysis one can either be interested in bad cells with abnormal indicator values in order to be able to fix them or in the behavior of the best cell in order to copy its configuration to other corresponding cells. In order to get correct information out of network data the used variables must be balanced by scaling. The most
Training Data set
SOM
Clustering
Classification Labeling
Fig. 2. Two phase clustering with classification.
In order to analyze a time-series data or a sequence of data over a time period instead of single data points, the frequency of appearance or the number of “hits” in each data cluster is computed for the given sequence of data. The vector containing these proportions or “hits” over some time period is called a hit-histogram. The hit-histogram of a cell over a time period provides the characterization of the cell behavior and is later used in clustering of the cells into similarly behaving groups. 3
The clustering of the cells is performed by processing the hit-histograms of the cells computed over consecutive time periods with a similar combination of the SOM and kmeans clustering algorithm (see Fig. 2) as in the extraction of the cell features (hit-histogram computation). At first, a SOM is trained using the hit-histogram vectors of each cell. Then, the codebook vectors of the SOM are clustered into different number of clusters. Finally, the best clustering is selected according to the Davies-Bouldin index. It should be noted here, that the hit-histogram vectors are considered as general feature vectors and the distance measure used in SOM training, SOM clustering and Davies-Bouldin index evaluation is the Euclidean distance measure.
variables are given in Table II. The frame error rate values are preprocessed using y = tanh(ax) function because it maps all x ≥ 0 into a range [0, 1] and the shape of the mapping can be controlled with the parameter a. In this way, the user has the possibility to focus on a certain range of FER values defined as interesting by the user, thus avoiding the possible dominance of the uninteresting phenomena in the data. TABLE II Key performance indicators
nUsr Number of users ulANR Uplink average noise raise ulFER Uplink frame error rate
D. Results of neural analysis The method described above has been used to analyze the uplink direction in the microcellular network scenario (see Fig. 5). This scenario was selected since it represents a challenging environment from the propagation point of view. Furthermore, the high capacity requirements of data services require a small cell environment. The WCDMA radio networks used in this study were planned to provide 64kbps service with 95% coverage probability, and with reasonable (2%) blocking. A Ray tracing model was used for the propagation loss estimation [14], [15]. The network layout comprises 46 omnidirectional base station sites. The selected antenna installation height was 10 meters in average. Due to the lack of measured data from live networks, simulated data produced by a dynamic system simulator [16] is used in the advanced analysis cases. During simulations, the multipath channel profile of the ITU Outdoor to Indoor A channel from [17] was assumed. The network parameters are collected in Table I. The system features used in the simulations are according to 3GPP [18]. The analysis results are presented in the following sections.
First, the structure of the data has been visualized using a SOM with 2D hexagonal grid of size 10×15. In Fig. 3 the component planes of the SOM are shown. Each component plane shows what kind of values a single variable has in different parts of the map. The value of the variable is indicated by gray-level and it can be read from the graylevel axis on the right side of the corresponding component plane. For example, the variable nUsr has values roughly between [0, 8] as can be seen from the gray-level axis on the right side of the component plane corresponding to the variable nUsr. The high values of nUsr are represented by dark gray-levels (as can be seen from the gray-level axis) and are located close to the lower right corner and the lower left corner of the map. Similarly, the top of the map represents data samples in which the values of nUsr are much lower (represented by light gray-levels). nUsr
7.69
TABLE I The network parameters during simulations.
Chip rate Base station (BS) maximum transmit power MS maximum transmit power MS minimum transmit power MS speed BS antennas MS antennas Propagation model Propagation channel profile
ulANR
20.7
3.87
3.84 Mchip/s 37 dBm
d
0.0364
ulFER
0.11
10.5
0.043
2.92
0.012
d
d
Fig. 3. SOM component planes.
21 dBm -44 dBm 3 km/h Omni, 11.0 dBi Omni, 0.0 dBi Ray tracing, in-building loss 12 dB Outdoor to Indoor A [17]
As mentioned earlier, the codebook vectors of the SOM have been clustered using the k-means algorithm in order to obtain a clustering for the data set. The properties of each data cluster can be analyzed using the component plane representation shown in Fig. 3 or using a set of automatically generated rules in order to find a quantitative description for the data clusters [19]. In Fig. 4, the clustered SOM (a) and the descriptive rules for the corresponding clusters are shown (b). The map consists of 7 clusters, each shown using a different gray-level. In addition, the map units are labeled according to the cluster in which the map units belong to. On the right are shown the descriptive rules that indicate the kind of data samples in the different clusters. From the rules it can be seen that, for
E. Uplink results in microcellular scenario For the analysis of the uplink direction in the microcellular scenario, three variables have been selected. Selected 4
example, data cluster 6 represents data samples with an unacceptably high ulFER. It should be noted here that the automatically generated rules represent the same information as the component plane representation shown in Fig. 3 although in quite different form.
(a)
(b)
(c)
(d)
tion into behavioral cluster 3. In Fig. 5b, the geographical locations of mobile cells and their corresponding classification results are shown. The dominant behavioral clusters in terms of the number of classified base stations are 1, 2 and 6. Behavioral clusters 1 and 2 are described by rules for data clusters 3 and 5, and for behavioral cluster 6 the rules for data clusters 2, 3 and 5 of Fig. 4b apply. Typical of these behavioral clusters is the relatively low number of users, good quality and low uplink noise raise value. The explanation for the low number of users is the relatively small cell dominance area, which is typical of microcellular networks. When comparing the geographical area of cells in behavioral clusters 1, 2 and 6, high correlation can be found with uplink loading. (See Fig. 7 produced by the traditional analysis. The light areas in the figure indicate low loading.) The only cell in the bad performance area, that is, the behavioral cluster 3 described by rules for data cluster 6 and 7, is cell 44. Characteristics of this cell are high load and high number of users. The FER performance of this cell is degraded and thus, it can be concluded that the cell is operating at the edge of its capability. It is worth noting that an analysis utilizing conventional means did not identify cell 44 as problematic (see Sec. IV-B). Only the use of SOM and further analysis using expert knowledge proved this. In Fig. 6, the trajectories of mobile cells 8, 14 and 44 that can be used for mobile cell monitoring purposes are shown. Cell 8 has been chosen as an example to demonstrate the behavior of an ”average”cell, which is surrounded from all directions with interfering cells (see Fig. 5b). Compared to the situation with cell 44, the difference is that this cell is better isolated from the surrounding cells, and thus, the performance tends to be better. Cell 44 is surrounded by water areas and interfering signals can freely propagate. Whereas in the case of Cell 8, the street canyon structure and buildings isolate Cell 8 from the interfering sources. Cell 8 starts from behavioral cluster 2 with almost all measured data points in data cluster 5 indicating very low ulANR and ulFER. Then the operation point moves into behavioral cluster 6 due to an increase in nUsr and ulANR, because the number of data samples in data clusters 2 and 3 has increased. Behavioral cluster 8 represents another increase in the number of data samples with the number of users up to 6-10 and ulANR between 2.4-9.1 dB. Then the operation point visits behavioral cluster 1 with data samples mostly in data clusters 3 and 5, which in turn indicates operation with low number of users. In the end, the operation point on the histogram map returns back to behavioral cluster 6. The behavior in this cell as a function of time changes rapidly, but the number of users and loading are strongly correlated, and thus, the interference from other cells does not dominate the loading. The loading is generated by the traffic in cell 8 itself, and thus, no capacity is wasted. Cell 14 operates mostly in behavioral clusters 1 and 2 with almost all of the data samples in data clusters 3 and 5. As mentioned earlier, these data clusters represent data samples that have low nUsr and ulANR. The behavior of
Fig. 4. (a) K-means clustering of the map of data vectors, (b) the single variable rules for data clusters and (c-d) K-means clustering of the histogram map.
In Fig. 4, the clusters of the histogram map (c) and the hit-histogram prototypes stored in each map unit of the histogram map are shown (d). The histogram map consists of 8 behavioral clusters (behavior characterized by a histogram), each indicated by a different gray-level. The map units describing the cell behavior are labeled according to the behavioral cluster in which they belong. From the figure it can be seen that behavioral cluster 3 on the histogram map consists of feature vectors in which most of the single data samples are located in data cluster 6. This can be seen from the sixth bar of the hit-histograms in the corresponding behavioral cluster. Since data cluster 6 represented data samples with high ulFER (see Fig. 4b), the behavioral cluster 3 characterizes a cell behavior as undesirable. In Fig. 5a, a classification of mobile cells using the histogram map for uplink direction data is shown. The classification is based on histogram features which are computed from a time window. From the figure it can be seen that mobile cell 44 has quality problems due to its classifica5
cells.
Cell 8
Cell 14
Cell 44
Fig. 6. Trajectories of the cells.
In Sec. IV-B, the cells with interference problems were identified by traditional means. These cells are 3, 6, 7, 18, 24, 28, 42 and 43. When checking the position of these cells in Fig. 5a, one can see that they are dominantly located in behavioral clusters 2, 4 and 7. Behavior in clusters 4 and 7 is rather similar as it is characterized with relatively high noise rise, but moderate number of users. This finding supports that the cause is the interference from other cells. Behavioral cluster 2 is characterized by a low number of users, and thus, the other cells’ interference is again dominating in the load factor.
(a) 1
1 1
1
8
7
2
6 7
6
6
8
1
1
2
6 1
1 6
8
4
A. Analytical approach for the network performance data
6
5
1
2
IV. Conventional analysis of WCDMA cellular network
1
1
6
6 4
4
6
1
2
7
2
6 4 1
1
2
3
7
As presented in [20], the uplink load factor ηUL can be calculated as a sum of load factors of all N uplink connections in a cell. The effect of the multicell environment is taken into account by multiplying the single cell case with the term (1 + i), where i is other to own cell interference ratio. Thus, the loading for a single service, fixed mobile station (MS) speed, multicell case is
7
6
6
(b) Fig. 5. (a) Mobile cell clustering and (b) locations of classified cells.
ηUL = N cell 14 can be explained with the geographical position of this cell. It is at the edge of the analyzed area and thus, it has less interfering neighbors than the other cells. Low level of interaction between the cells and low number of users explain the rather static behavior of this cell. The behavioral clusters in question are characterized as ones having relatively low loading and good quality performance. Cell 44, located on an island off the coast of the city, first operates in behavioral clusters 7 and 8 with data samples in data clusters 1, 2, 4 and 7. This is the lower left corner of the map in Fig. 4b which represents data samples with different loads but low ulFER. However, the operation point moves into behavioral cluster 3 due to an increase in the number of samples in data cluster 6, that is, the cluster with the highest values for ulFER. Characteristic of this cell is that the loading ranges (both uplink and downlink) from moderate to high, but the number of users remains insignificant. The low number of users and high loading indicates interference problems due to interference from other
1 (1 + i) W 1 + ρR
(4)
in which R is the used bit rate, ρ is the signal-to-noise ratio requirement and W is the WCDMA chip rate. B. Traditional analysis results for microcellular case Using Eq.(4), two reference figures for the cell performance can be calculated based on the input data, namely the loading caused by a user and the number of users a cell can serve: The used uplink Eb /No value in this study was 3.5 dB. During the simulations loading was set to 0.95. Frequently quoted i value is 55%. The theoretical capacity values are presented in Table III. In addition to these analytical values the network was analyzed with a static radio network planning tool. These results from the radio network planning tool and especially the format used in visualization of the results is what the network management systems at their best can offer today: visualization support of individual KPIs and textual reports containing measurements and alarms. 6
TABLE III Theoretical capacity values.
Number of users, upper bound Loading/user, upper bound Number of users, i included Loading/user, i included
seen that the situation is not straightforward. Typically, admission control limits the number of users if the loading caused by active users is high. In this case, loading seems not to be the reason for blocking. Blocking is caused by downlink power outage, i.e. the maximum power allowed for one individual user. This explanation is based on the fact that microcellular scenarios are often downlink limited. In Table IV example results for the microcellular case are introduced. Only some example cells are chosen, the cells are the same as in the SOM trajectory analysis, i.e., cells 8, 14, 44. Full set of results can be found in Appendix in Table V.
26 0.03621 16 0.05613
BS14
BS1 BS13
BS2
BS15 BS3
BS35
BS4
BS16
BS24
BS34
BS39
BS32
BS23 BS17
BS10
BS40
BS33
BS26 BS11 BS5
TABLE IV Example results from traditional analysis.
BS25
BS12
Microcells
BS36
BS38
BS27
BS18
BS31
BS37 BS41
BS6
BS9
BS22
BS28
BS19
BS30
BS20
BS21
Cell Id BS Txp Traffic [W] Loading Other to own cell interference ratio, i Users UL Throughput UL [kbit/s]
BS44 BS46
BS43
BS8 BS7
BS42
BS45
BS29
(a)
BS15 BS25
BS12
BS35
BS4
BS16
BS24 BS26
BS34
BS17
BS10
BS36 BS27
BS18 BS6
BS39
BS32
BS23 BS5
BS40
BS33
BS11
BS31
BS38
BS37 BS41
BS9
BS22
BS28
BS19
BS30
BS20
BS21
BS42
BS44 BS46
BS43
BS8 BS7
BS45
BS29
22 18 22 1408 1152 1408
In Eq.(5), ThroughputN ORM is the normalized throughput. In the normalization, the maximum throughput was the maximum value in a cell. According to this classification the top 10% performing cells were cells 8, 9, 11, 25, 29 and 44.
Cell loading %
0.55
44 0.39 0.88 0.15
All of the selected microcells have relatively high loading, and significantly low i, compared to the often used value 55%. This indicates very good isolation of the cells. For all of these cells the loading per user is 0.04, which is very close to the upper bound value in Table III. In general, all the microcells have a well-controlled interference situation, only 8 cells out of 46 had i higher than 55%. These cells are: 3, 6, 7, 18, 24, 28, 42, and 43. For uplink performance PUL evaluation a simple function f combining the interference control and throughput aspects were generated, see Eq.(5). The weighting for each item in the cost function was the same. PUL = f (ThroughputN ORM , i, ηUL ) (5)
BS13
BS3
0.5
14 0.25 0.75 0.13
BS14
BS1 BS2
8 0.40 0.88 0.16
0.6
0.65
0.7
0.75
0.8
0.85
0.9
0.95
(b) Fig. 7. (a) Used network scenario and (b) Uplink loading. Site distance is couple of hundreds of meters. In (a) users with transmission power problems are denoted with ⋄ and ⋆ for uplink and downlink respectively. In (b) the lighter the gray-level the lower is the uplink loading.
V. Evaluation of the SOM based method A. Validity of the SOM results During analytical analysis (Sec. IV-B) for the microcell uplink case, the best performing cells were 8, 9, 11, 25, 29 and 44. When mapping these cells on the performance spectrum of Fig. 5a it is very interesting to note that cells 8, 9, 11 and 29 are all in the same behavioral cluster (i.e. 6). Cell 25 is in behavioral cluster 2, owing to the fact that as an edge cell (dominance area mostly on water) it has a lot less users than the other cells. Traditional means are not capable of finding the performance problems of cell 44. As noted earlier in Sec. III, this cell is performing at
In Fig. 7b, the uplink loading is visualized. The darker the gray-level the higher is the loading in a cell. In Fig. 7a, the location of the mobile stations that suffer from power outage are depicted. When uplink and downlink information is compared, it can be noted that the locations tend to correlate. Thus, uplink performance and downlink performance are well in balance. When the correlation between loading and the number of uplink users is tested it can be 7
the edge of its capabilities. Similar results can be found in the macrocellular and microcellular case when both the uplink and the downlink direction are analyzed [21]. As a conclusion, it can be said that traditional means support the conclusions of the cell classification performed by the SOM.
solutions that actually disturb the overall quality of the cell.
B. Convenience and usability of the SOM based analysis Current network performance monitoring and analysis tools are not capable of meeting the needs and requirements of service driven networks. The reason for this is the increasing number of measurements that one should process simultaneously. The introduction of each new service or service class will increase the amount of measurements the operator needs to collect from the network elements. Fig. 8 represents an example of a typical KPI output. Each measurement is presented separately and the end user is responsible for correlating the different measurements. Furthermore, there is also a physical limitation as to how many measurements can be visualized at once. The end user is responsible for combining the information from different domains.
Fig. 9. Measurement example from NMS with different time resolutions. For the data a minute, an hour and daily average filter has been used.
Deriving the QoE from solely objective measurements requires advanced intelligence and tools, like the SOM. The method proposed in this paper is highly visual and it can be used to combine different types of information like time and geography with the cell behavior profile. This is different from the traditional analysis where only a few KPIs are analyzed separately at a time. C. Scalability of the method As mentioned earlier, the network performance is currently handled by visualizing the KPIs separately on the screen. As a result, the number of graphs to be visualized increases w.r.t the number of services, the number of network elements and the number of KPIs used in the analysis. Thus, the optimization of the total performance over a large set of network elements, covering also all of the most important quality aspects, requires the analysis of a large number of separate graphs, and the combination of their results to form a single big picture of the current state of the network. However, the number of required visualizations does not increase so radically in the SOM-based approach presented in this paper. For example, the increase in the number of network elements increases only the number of cell trajectories (as in Fig. 6). If the number of services or the number of network measurements (variables) increases, the number of component plane visualizations as in Fig. 3 also increases. If only the most descriptive combination of the single variable rules for each data cluster is shown, the number of symbolic rules per data cluster increases with the number of used variables only if the new set of variables is able to separate the data clusters better than the old set of variables. Thus, the number of rules per cluster does not necessarily grow at the same rate as the number of variables used in the analysis. The SOM-based method is more scalable with respect to the changes in the problem setting than the currently used methods.
Fig. 8. An example of measurement visualization in network management system. For each measurement own window is needed.
Current methods rely strongly on averaging or cumulating values over longer period of time, most often one day. KPI values are analyzed as snapshots representing one period. This approach loses details such as the form of value distribution of a KPI inside the period. The approach can be enhanced by dividing the period into sub-periods and calculating averages or cumulative sums over those. This easily generates an amount of data so large that details vanish into them and the simultaneous analysis of KPI combinations or performance of cell group becomes even harder. Fig. 9 presents an example of NMS level KPI visualization. The same KPIs are presented with minute, hour and daily average filters. When a problematic KPI is detected, and it is further analyzed, it can happen that the actual behavior pattern of the cell and its development is not detected at all but the operator ends up optimizing or fixing only one aspect of the cell’s behavior. This can lead to sub-optimal 8
D. Computational complexity of the method The method consists of two independent algorithms (SOM, k-means) whose computational complexity is dependent on the amount of analyzed data. According to [10], the computational complexity of the SOM is proportional to N M d, where N is the number of data samples in the data set, M is the number of map units in the SOM and d is the number of variables (dimension) of the data set. Thus, the training of relatively small SOMs (less than thousand map units) is computationally efficient even for large data sets. In [13], the computational complexity of the k-means algorithm and its application to the clustering of the SOM codebook vectors is described. The computational complexity of the basic P k-means clustering algorithm is said to max be proportional to C k=2 N k, where N is the number of data samples and Cmax is the maximum number of clusters in which the data is to be clustered. However, when the kmeans clustering is applied to the codebook vectors of the SOM trained with the data set instead of the original data set, the computational PCmax cost of clustering is reported to reduce to N M + k=2 M k, thus making it computationally tractable solution for the analysis process described in this paper. As a reference, the computation time for a single network scenario was less than an hour, including the basic computation, visualization and report generation. VI. Usage of clustering in optimization The analysis method presented and used in this article consists of two different phases: the clustering of single data points consisting of several KPIs and clustering of sequences of these data points. The first phase is actually the traditional way to use the SOM. It is able to show what types of KPI value combinations there are in general. The basic clustering can be analyzed further in order to see how an average behavior of cells has developed during a longer period of time. In the case of behavioral clusters, the method takes into account not only the average behavior pattern based on all selected KPIs but also the recent variation in behavior pattern. By doing this, the method is able to detect the collapse of an excellent behavior much earlier than by analyzing only the average behavior. Cell clusters that have been found by the proposed methods can be used as a starting point for a more detailed analysis. This is especially suitable for trouble shooting type of tasks. For example, in Fig. 10, it can be assumed that the lower left corner of the SOM in the upper figure indicates cells that have a certain defect. These cells can automatically be selected as optimization targets. More detailed analysis can be provided on them and probably also actions to fix them can be suggested. The ability to combine the result provided by the SOM and the geographical locations of the cells is essential for an operator. The geographical location information can provide support to problem resolution. The cell clusters can also be used to simplify the task of parameter provisioning and optimizing. It is not feasible technically nor timewise to optimize the network cell by cell. Cell clustering can be used to increase the operational efficiency of an operator’s tasks. The cells can be
Fig. 10. Example of utilization of cluster information in optimization process. Selection for cells to be optimized/autotuned.
clustered based on their traffic profiles and density, propagation conditions, cell types, radio resource management functionality performance etc. When homogeneous meaningful groups have been found, all the cells in one group can share a set of configuration parameter values. This kind of grouping based on multiple criteria is more accurate and the operation of the network will benefit from this: the optimization process is greatly simplified, improved and made less error prone and more efficient. In behavioral clustering, the performance change over time can be shown on the SOM (see Fig. 6). The method actually combines two sources of information: change over time and type of behavior at each separated point of time. This type of analysis can be performed using data averaged over various time periods, ranging from tens of seconds to days. One could, for example, follow one cell movement in the SOM during peak traffic hours, assuming that networks are able to report cell performance frequently enough. The sequence clustering method can be applied to follow up the optimization actions made in the network. When the network element configuration is changed, the operator normally wishes to see the effect of the change to performance. The method can be used to show how the changed cells change their places on the SOM. This kind of infor9
mation combination will be essential in the optimization of WCDMA radio networks. The complexity of the radio networks is growing as well as the size of networks themselves. Operators will need means to rapidly analyze the changes in the network, given the high number of cells, several services with different QoS criteria and a large amount of collected performance data.
Acknowledgment The authors wish to thank Albert H¨oglund, Jukka Henriksson, Ari H¨am¨al¨ainen, Mikko T. Toivonen, Hannu Multim¨aki and other colleagues from Nokia and Sampsa Laine from Helsinki University of Technology for their valuable comments. References
VII. Conclusion
[1]
This paper proposes the use of the Self-Organizing Maps (SOM), a neural network method, in the analysis phase of high level telecommunication network optimization process. The results provided by the SOM were verified with the combination of traditional means and expert knowledge. It can be stated that the results show a good agreement. Being able to understand the SOM results with traditional means increases confidence in the novel analysis and its applicability in the area of cellular networks. During the course of this work it was noticed that the traditional analysis as such is not adequate enough to provide as enhanced demonstration of the network performance as the SOM provides. Current performance analysis methods include combinations of planning tool analysis data and real network measurements. The information format in a planning tool is a static snapshot of the network situation, and thus, the aspect of time is not present. The KPIs offered by the NMS are averaged values. Certain amounts of information of the past behavior of the KPI is also available. In the case of KPI information, no prior knowledge on the correlation of the different KPIs is available. The correlation analysis and the combination of these two information sources is done manually and it strongly relies on the expertise of the person performing the task. The strength of the SOM based analysis methods is in the fact that multiple measurements are used in the analysis at the same time. Furthermore, the output can be provided in a descriptive format to ease the operator decisions. The SOM based application in the analysis of cellular networks is not widely spread, and it is worth noting that the SOM based analysis has been successfully applied to GSM data to detect anomalous behavior of base stations, see [22] for more details. In the future, the operation of cellular networks will be strongly service driven. Compared to the current situation with provisioning of voice and simple best effort data services only, the change is enormous. Effective analysis of 2G networks’ voice service is currently challenging enough because of large volumes of data collected from network elements. The evolution towards 3G systems will further increase the amount of data. The operators’ task is to filter the relevant information to a level on which it can be easily handled. Furthermore, the data set must include all the essential parts needed to conclude the service quality. Operators can benefit from the introduced neural methods already today. The full gain and potential of the advanced analysis methods can be reached when multiple end-user services are provided, and the quality perceived by the customers need to be monitored and optimized.
[2] [3]
[4]
[5] [6]
[7] [8]
[9]
[10]
[11] [12] [13] [14] [15] [16]
[17] [18] [19]
[20] [21]
10
Telemanagement Forum, “Enhanced telecom operations map eTOM, The Business Process Framework for the Information and Communications Services Industry, Version 3.0,” Tech. Rep. GB921, June 2002. T. Kohonen, Self-Organizing Maps, Springer-Verlag, Berlin, 1995. K. Raivio, O. Simula, and J. Laiho, “Neural analysis of mobile radio access network,” in IEEE International Conference on Data Mining, San Jose, California, USA, November 29 - December 2 2001, pp. 457–464. P. Lehtim¨ aki, K. Raivio, and O. Simula, “Mobile radio access network monitoring using the self-organizing map,” in European Symposium on Artificial Neural Networks, Bruges, Belgium, April 24 - 26 2002, pp. 231–236. T. Kohonen, E. Oja, O. Simula, A. Visa, and J. Kangas, “Engineering applications of the self-organizing map,” Proceedings of the IEEE, vol. 84, no. 10, pp. 1358–1384, October 1996. E. Alhoniemi, J. Hollm´ en, O. Simula, and J. Vesanto, “Process monitoring and modeling using the self-organizing map,” Integrated Computer Aided Engineering, vol. 6, no. 1, pp. 3–14, 1999. O. Simula, P. Vasara, J. Vesanto, and R.-R. Helminen, Industrial Applications of Neural Networks, chapter The Self-Organizing Map in Industry Analysis, pp. 87–112, CRC Press, 1999. O. Simula, J. Ahola, E. Alhoniemi, J. Himberg, and J. Vesanto, “Self-organizing map in analysis of large-scale industrial systems,” in Proceedings of the Workshop on Self-Organizing Maps, Espoo, Finland, July 1 - 3 1999, pp. 375–387, (invited paper). T. Kohonen, J. Hynninen, J. Kangas, J. Laaksonen, and K. Torkkola, “SOM PAK: The self-organizing map program package,” Tech. Rep. A31, Helsinki University of Technology, Laboratory of Computer and Information Science, 1996, Available: http://www.cis.hut.fi/research/som pak/. J. Vesanto, J. Himberg, E. Alhoniemi, and J. Parhankangas, “SOM toolbox for Matlab 5,” Tech. Rep. A57, Helsinki University of Technology, Laboratory of Computer and Information Science, 2000, Available: http://www.cis.hut.fi/projects/somtoolbox/. B.S. Everitt, Cluster Analysis, Arnold, 1993. D.L. Davies and D.W. Bouldin, “A cluster separation measure,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 1, no. 2, pp. 224–227, April 1979. J. Vesanto and E. Alhoniemi, “Clustering of the self-organizing map,” IEEE Transactions on Neural Networks, vol. 11, no. 3, pp. 586–600, May 2000. K. Heiska and A. Kangas, “Microcell propagation model for network planning,” in Personal, Indoor and Mobile Radio Communications, 1996, vol. 1, pp. 148–152. J. Rajala, K. Sipil¨ a, and K. Heiska, “Predicting in-building coverage for microcells and small macrocells,” in IEEE Vehicular Technology Conference, 1999, vol. 1, pp. 180–184. S. H¨ am¨ al¨ ainen, H. Holma, and K. Sipil¨ a, “Advanced WCDMA radio network simulator,” in Personal, Indoor and Mobile Radio Communications, Osaka, Japan, September 12-15 1999, vol. 2, pp. 951–955. “Guidelines for evaluation of radio transmission technologies for IMT-2000,” 1997, Recommendation ITU-R M. 1225. “http://www.3gpp.org/,” . M. Siponen, J. Vesanto, O. Simula, and P. Vasara, “An approach to automated interpretation of SOM,” in Advances in Self-Organizing Maps, N. Allinson, H. Yin, L. Allinson, and J. Slack, Eds. 2001, pp. 89–94, Springer. J. Laiho, A. Wacker, and T. Novosad, Eds., Radio Network Planning and Optimisation for UMTS, John Wiley & Sons Ltd., 2001. J. Laiho, K. Raivio, P. Lehtim¨ aki, K. H¨ at¨ onen, and O. Simula, “Advanced analysis methods for 3G cellular networks,”
Tech. Rep. A65, Helsinki University of Technology, Laboratory of Computer and Information Science, 2002. [22] A.J. H¨ oglund, K. H¨ at¨ onen, and A.S. Sorvari, “A computer hostbased user anomaly detection system using the self-organizing map,” in Proceedings of the International Joint Conference on Neural Networks, 2000, vol. 5, pp. 411–416.
Appendix
11
TABLE V Results from traditional analysis Cell Id BS Txp Traffic [W] Loading Other to own cell interf. ratio, i Own cell loading ∆ loading caused by other cells [%] Users UL Loading/user Own cell interf./user Throughput UL [kbit/s] Cell Id BS Txp Traffic [W] Loading Other to own cell interf. ratio, i Own cell loading ∆ loading caused by other cells [%] Users UL Loading/user Own cell interf./user Throughput UL [kbit/s] Cell Id BS Txp Traffic [W] Loading Other to own cell interf. ratio, i Own cell loading ∆ loading caused by other cells [%] Users UL Loading/user Own cell interf./user Throughput UL [kbit/s]
1 0.39
2 0.24
3 0.46
4 0.56
5 0.28
6 0.15
7 0.15
8 0.40
9 0.41
10 0.23
11 0.28
12 0.46
13 0.32
14 0.25
15 0.30
16 0.24
0.84 0.32
0.72 0.51
0.89 0.88
0.89 0.20
0.70 0.28
0.65 0.65
0.65 0.59
0.88 0.16
0.89 0.07
0.89 0.29
0.89 0.12
0.89 0.53
0.89 0.28
0.75 0.13
0.89 0.40
0.73 0.32
0.63 24
0.47 34
0.47 47
0.74 17
0.55 22
0.39 40
0.41 37
0.76 14
0.83 7
0.69 23
0.79 11
0.58 35
0.69 22
0.66 12
0.64 29
0.55 24
19 12 16 22 15 11 12 22 25 21 23 17 21 18 19 15 0.044 0.060 0.055 0.040 0.047 0.059 0.054 0.040 0.035 0.042 0.039 0.053 0.042 0.042 0.047 0.049 0.033 0.040 0.029 0.033 0.036 0.035 0.034 0.035 0.033 0.033 0.034 0.034 0.033 0.037 0.033 0.037 1216
768
1024
1408
960
704
768
1408
1600
1344
1472
1088
1344
1152
1216
960
17 0.32
18 0.14
19 0.40
20 0.38
21 0.40
22 0.24
23 0.39
24 0.33
25 0.51
26 0.21
27 0.19
28 0.11
29 0.37
30 0.49
31 0.43
32 0.14
0.89 0.23
0.70 0.68
0.89 0.18
0.74 0.27
0.76 0.21
0.85 0.36
0.88 0.19
0.89 1.20
0.87 0.13
0.72 0.39
0.81 0.28
0.60 0.89
0.86 0.12
0.87 0.28
0.88 0.27
0.52 0.46
0.73 18
0.42 40
0.76 15
0.58 21
0.63 18
0.62 27
0.74 16
0.40 55
0.77 12
0.52 28
0.63 22
0.32 47
0.77 11
0.68 22
0.69 21
0.36 31
22 12 22 16 18 18 21 16 22 15 18 9 23 20 21 10 0.040 0.058 0.041 0.046 0.042 0.047 0.042 0.056 0.039 0.048 0.045 0.067 0.038 0.044 0.042 0.052 0.033 0.035 0.034 0.037 0.035 0.035 0.035 0.025 0.035 0.035 0.035 0.035 0.033 0.034 0.033 0.036 1408
768
1408
1024
1152
1152
1344
1024
1408
960
1152
576
1472
1280
33 0.38
34 0.22
35 0.23
36 0.28
37 0.34
38 0.36
39 0.22
40 0.17
41 0.35
42 0.12
43 0.16
44 0.39
45 0.29
46 0.30
0.72 0.31
0.80 0.26
0.79 0.20
0.89 0.28
0.79 0.53
0.87 0.37
0.68 0.31
0.64 0.51
0.89 0.23
0.46 0.85
0.75 0.63
0.88 0.15
0.77 0.30
0.85 0.21
0.56 23
0.63 20
0.66 17
0.69 22
0.51 35
0.63 27
0.52 24
0.42 34
0.73 19
0.25 46
0.46 39
0.76 13
0.59 23
0.70 17
14 18 19 20 15 18 12 12 22 6 13 22 16 21 0.052 0.044 0.042 0.044 0.052 0.048 0.056 0.053 0.041 0.076 0.057 0.040 0.048 0.040 0.040 0.035 0.035 0.035 0.034 0.035 0.043 0.035 0.033 0.041 0.035 0.035 0.037 0.033 896
1152
1216
1280
960
1152
768
12
768
1408
384
832
1408
1024
1344
1344
640
I. Introduction The mobile communication industry is currently shifting its focus from second generation networks (2G) towards the third generation networks (3G). The shift is not only related to the evolution of the access technology, but also to the vision of the development of service provisioning and service demands, customer expectations and customer differentiation. While current wireless networks still evolve and service providers bring new internet packet data services into the markets, an increasing number of operators and other wireless communication professionals are becoming familiar with the wideband code division multiple access (WCDMA) technology and prepare themselves for 3G services and networks. There will be a number of new challenges when shifting from the current 2G to the new 3G networks, many of them related to the design and the operation of true multi-service radio networks. An essential part of the new challenges is related to the provisioning, monitoring and optimization of the services. The number of network counters and measurements shall increase owing to the fact that instead of monitoring GSM voice only, one must concentrate on monitoring multi system and multiservice environment. The information space of future netJ. Laiho and K. H¨ at¨ onen are with Nokia Group, Finland K. Raivio, P. Lehtim¨ aki and O. Simula are with the Neural Networks Research Centre, Helsinki University of Technology, Espoo, Finland The study has been financed by Nokia Networks, Nokia Mobile Phones and National Technology Agency of Finland (TEKES) which is gratefully acknowledged.
works is manifold compared to that of today’s networks. This fact is the main driver pushing development towards advanced network analysis solutions, like the one presented in this paper. Such functionality is located in the network and service management layers of the Telecom Operations Map (TOM) framework [1]. Operating a cellular network is an iterative quality cycle process combining the network and service configuration and the related performance measurements. In this cycle, the overall end-to-end quality target is defined and the quality criteria and thresholds for key performance indicators (KPI) for each service type are determined. Network performance data is gathered from Network Management Systems (NMS), drive tests, protocol analyzers and/or customer complaints. The actual measured service performance is analyzed and the results are compared with the set targets. In case of conflict, corrective actions to the network configuration are carried out. Effective analysis methods are prerequisites for dynamic and successful operations. This paper concentrates on the analysis and visualization part of the quality cycle. Datadriven algorithms and data mining methods provide efficient tools for exploratory data analysis. Algorithms based on Artificial Neural Networks (ANNs) have proved to be especially suitable in highly complex and data intensive applications. The Self-Organizing Map [2] is one of the most popular neural algorithms due to its efficient visualization properties. The motivation for the introduction of neural analysis on the network performance data is to provide effective means to handle multiple KPIs simultaneously. Furthermore, effective analysis methods reduce operators’ trouble shooting efforts, speed up the cycle, and thus, the network utilization rate increase. Furthermore, the quality as experienced by the end user (QoE) becomes increasingly important owing to the fact that over provisioning quality is inefficient and expensive. On the other hand, customers suffering from low quality tend to change to the competitor’s network. Determining the QoE is about collection and combination of information from different domains (access networks, interfaces and core network etc.). In this paper, an application of the SOM in analyzing telecommunications networks is presented. Mobile cells of a WCDMA network are classified according to their performance [3], [4]. SOM applications in a cellular environment are new and thus the results are verified using analytical results and expert knowledge. The Self-Organizing Map is introduced in Section II. Radio access network analysis methods based on the SOM and the results of the analysis are presented in Section III. In Section IV, the same
network is analyzed using conventional methods. Methods and results of the novel SOM based analysis are evaluated in Section V. Finally, the usability of the new method is discussed in Section VI.
data samples and prototype vectors in the input space. The update step can be performed by applying
II. The Self-Organizing Map
where α(t) is the learning rate and hc (t, i) is the neighborhood function of the algorithm. The last term in the square brackets is proportional to the gradient of the squared Euclidean distance d(x, mi ) = ||x − mc ||2 . The learning rate α(t) ∈ [0, 1] is usually a monotonically decreasing function of time. A good candidate is α(t) = α0 (1 − t/T ), where α0 is the initial value for the learning rate and T is the total number of training iterations. A very frequently used form for the neighborhood function hc (t, i) is the Gaussian one centered on the winner map unit c: ||ri − rc ||2 (3) hc (t, i) = exp − 2σ(t)2
mi (t + 1) = mi (t) + α(t)hc (t, i)[x(t) − mi (t)]
A. SOM in data mining Data mining and exploration is an emerging area of research in artificial intelligence and information management. The objective of data mining is to extract relevant information from large amounts of data. Data mining and analysis tasks typically include clustering, classification, and regression of data, determining parameter dependencies, and finding various anomalies from the data. Artificial Neural Networks provide efficient tools in data mining and they have been successfully used in various data engineering applications. The Self-Organizing Map (SOM) is a widely used neural network algorithm [2]. It maps a high-dimensional data manifold onto a low-dimensional, usually two-dimensional, grid or display. The SOM has several beneficial features which make it a useful tool in data mining and exploration. The SOM follows the probability density function of the data and is, thus, an efficient clustering and quantization algorithm. However, the most important feature of the SOM in data mining is the visualization property. The topology preserving property of the SOM mapping results in a display inherently visualizing the clusters in the data. SOM based methods have been applied in the analysis of process data, for example, in steel and forest industry [5], [6], [7], [8].
(2)
where rc depicts the coordinates of the winner unit c, and ri denotes the coordinates of an arbitrary unit i on the discrete output lattice of the map and σ(t) is the width of the neighborhood. It is necessary that hc (t, i) → 0 when t → ∞ for the algorithm to converge. During learning, the learning rate and the width of the neighborhood function are decreased, typically in a linear fashion. The map converges to a stationary distribution, which approximates the density of the data. One step of the training algorithm of the SOM is illustrated in Fig. 1. The size of the SOM is 16 units, which have been arranged into a two-dimensional grid of 4 by 4 units. A data sample is marked with a cross; the black circles are the values of the prototype vectors before, and the gray circles after updating them towards the data sample. This kind of an update step is repeated iteratively during the training process. There exist two freely available software packages that include implementation of the SOM: SOM-PAK [9] and SOM Toolbox for Matlab [10].
B. SOM algorithm The basic SOM consists of a regular grid of map units or neurons. They are connected to the neighboring units using, for instance, a rectangular or hexagonal neighborhood. Each map unit, denoted here by i, is represented by a prototype vector mi . The dimension of prototype vectors is equal to the dimension of the input data. First, the prototype vectors are initialized, for instance, to random values. Then, during the training the values of the prototype vectors are adapted to follow the properties of the input data. Training of the SOM is divided into two alternating steps, typically thousands of times each. First, one data vector x from the training data set is randomly selected and the corresponding best-matching (winner) unit (BMU) c is determined. The prototype vector of the BMU, denoted by mc , is the one that is nearest to the data sample. In other words, it minimizes the Euclidean distance between x and mc : c = arg min ||x − mi || (1)
Fig. 1. An illustration of the SOM training.
i
C. SOM in Network Analysis
In the next step, the prototype vectors of the winner and its neighbors are moved towards the data vector. It should be noted that the neighborhood is defined in terms of the lattice structure, not according to the distances between
A behavior pattern of a cell at a certain instant is a set of indicator values that have been recorded at that instant. In network analysis, the SOM can be used to find 2
common method to do the balancing is to normalize the variance of each variable equal to one. After the normalization, the distribution of the data might be skewed if there are outliers in the data. If the average normal behavior is studied, the usual solution is to remove outliers or to replace them with estimated normal or correct values. If outliers carry interesting information, for example, as is the case in our study, where they can be signs of network problems that are searched for, it is possible to keep outliers but prevent their large values from dominating the analysis results. This can be done by using some sort of conversion function like tanh(x) (or log(x)) before the variance normalization. Such a function can decrease the effect of outliers and emphasize proper parts of the distribution. For example, the tanh function emphasizes small values at the expense of large values.
and show similarities between behavior patterns of cells. These indicators can include, for example, any subset of those indicators that are used in the traditional analysis of a network. Indicators that have been used in this paper, are shown in Table II. A set of n indicator values form an n-dimensional pattern vector. During the training phase a set of these vectors is used to train a SOM. During the training, a SOM approximates the distribution of pattern vectors in the n-dimensional space so that everywhere in the space, where there are vectors, there are nodes of a SOM map as well. A prototype vector can be a BMU for several data samples that are almost similar. The SOM gives a topology preserving mapping of the input. SOM nodes that are direct neighbors in a SOM grid pull themselves towards each other. So when the training has been done, most of the SOM nodes neighboring each other in the grid are placed next to each other also in the input space. Therefore, cell behavior patterns that have been mapped to neighboring SOM nodes usually resemble each other. When the SOM is visualized, similarly behaving cells can be spotted close to each other.
C. Clustering analysis In general, clustering is the grouping of similar samples together. In this work, clustering is also used to find groups of similarly behaving cells. The features used in mobile cell clustering and classification are computed with the method shown in Fig. 2. The data vectors of all the cells are clustered using a combination of the Self-Organizing Map and the k-means algorithm. At first, a SOM with M map units is trained using the data vectors. Next, the set of M codebook vectors of the SOM are clustered into several different numbers of clusters using the k-means algorithm [11]. The clustering process can be repeated several times for different values of k, for √ example, 100 clusterings for each value of k, 2 ≤ k ≤ M . The k-means clustering has to be run several times for each k because the algorithm gives different results depending on the initialization. The best clustering for each k minimizes the sum of squared errors. Several values of k have to be tested because the correct number of clusters is not known. The optimal value of k is defined using some clustering validity index, like the Davies-Bouldin index [12]. Such an index is able to tell us the best number of clusters for the current data set. In this algorithm, the SOM is used to quantize the data and to visualize the cluster structure in the data because it is much faster to find the clusters of SOM codebook vectors than the clusters of original data directly [13].
III. Network analysis using SOM In this section, a neural method for classification of mobile cells is presented. The method consists of the following phases: target selection, data preprocessing, clustering analysis and result interpretation. A. Target selection The first step in the process is the target definition. This includes the selection of the geographical area, network objects (base stations, radio network controllers, routers, etc.) and visualization task specification. The selection of network objects and the visualization task have a strong impact on the selection of the measurements and KPIs to be analyzed. Naturally each object in the network has its own specific measurements. The visualization task can be more generic or problem oriented. General performance analysis requires a different set of measurements than a specific trouble shooting case. B. Preprocessing Neural network methods are multivariate methods that study the combination of variables, i.e., their joint distribution. Before they can be applied to the data, the data has to be prepared for the analysis in a preprocessing phase. The main objective of the preprocessing phase is to ensure that the analysis methods are able to extract correct and needed information from the data. Preprocessing can filter out noise, handle the problem of missing values and balance different variables and their value ranges. The action required stems from the current information need. For example, in network analysis one can either be interested in bad cells with abnormal indicator values in order to be able to fix them or in the behavior of the best cell in order to copy its configuration to other corresponding cells. In order to get correct information out of network data the used variables must be balanced by scaling. The most
Training Data set
SOM
Clustering
Classification Labeling
Fig. 2. Two phase clustering with classification.
In order to analyze a time-series data or a sequence of data over a time period instead of single data points, the frequency of appearance or the number of “hits” in each data cluster is computed for the given sequence of data. The vector containing these proportions or “hits” over some time period is called a hit-histogram. The hit-histogram of a cell over a time period provides the characterization of the cell behavior and is later used in clustering of the cells into similarly behaving groups. 3
The clustering of the cells is performed by processing the hit-histograms of the cells computed over consecutive time periods with a similar combination of the SOM and kmeans clustering algorithm (see Fig. 2) as in the extraction of the cell features (hit-histogram computation). At first, a SOM is trained using the hit-histogram vectors of each cell. Then, the codebook vectors of the SOM are clustered into different number of clusters. Finally, the best clustering is selected according to the Davies-Bouldin index. It should be noted here, that the hit-histogram vectors are considered as general feature vectors and the distance measure used in SOM training, SOM clustering and Davies-Bouldin index evaluation is the Euclidean distance measure.
variables are given in Table II. The frame error rate values are preprocessed using y = tanh(ax) function because it maps all x ≥ 0 into a range [0, 1] and the shape of the mapping can be controlled with the parameter a. In this way, the user has the possibility to focus on a certain range of FER values defined as interesting by the user, thus avoiding the possible dominance of the uninteresting phenomena in the data. TABLE II Key performance indicators
nUsr Number of users ulANR Uplink average noise raise ulFER Uplink frame error rate
D. Results of neural analysis The method described above has been used to analyze the uplink direction in the microcellular network scenario (see Fig. 5). This scenario was selected since it represents a challenging environment from the propagation point of view. Furthermore, the high capacity requirements of data services require a small cell environment. The WCDMA radio networks used in this study were planned to provide 64kbps service with 95% coverage probability, and with reasonable (2%) blocking. A Ray tracing model was used for the propagation loss estimation [14], [15]. The network layout comprises 46 omnidirectional base station sites. The selected antenna installation height was 10 meters in average. Due to the lack of measured data from live networks, simulated data produced by a dynamic system simulator [16] is used in the advanced analysis cases. During simulations, the multipath channel profile of the ITU Outdoor to Indoor A channel from [17] was assumed. The network parameters are collected in Table I. The system features used in the simulations are according to 3GPP [18]. The analysis results are presented in the following sections.
First, the structure of the data has been visualized using a SOM with 2D hexagonal grid of size 10×15. In Fig. 3 the component planes of the SOM are shown. Each component plane shows what kind of values a single variable has in different parts of the map. The value of the variable is indicated by gray-level and it can be read from the graylevel axis on the right side of the corresponding component plane. For example, the variable nUsr has values roughly between [0, 8] as can be seen from the gray-level axis on the right side of the component plane corresponding to the variable nUsr. The high values of nUsr are represented by dark gray-levels (as can be seen from the gray-level axis) and are located close to the lower right corner and the lower left corner of the map. Similarly, the top of the map represents data samples in which the values of nUsr are much lower (represented by light gray-levels). nUsr
7.69
TABLE I The network parameters during simulations.
Chip rate Base station (BS) maximum transmit power MS maximum transmit power MS minimum transmit power MS speed BS antennas MS antennas Propagation model Propagation channel profile
ulANR
20.7
3.87
3.84 Mchip/s 37 dBm
d
0.0364
ulFER
0.11
10.5
0.043
2.92
0.012
d
d
Fig. 3. SOM component planes.
21 dBm -44 dBm 3 km/h Omni, 11.0 dBi Omni, 0.0 dBi Ray tracing, in-building loss 12 dB Outdoor to Indoor A [17]
As mentioned earlier, the codebook vectors of the SOM have been clustered using the k-means algorithm in order to obtain a clustering for the data set. The properties of each data cluster can be analyzed using the component plane representation shown in Fig. 3 or using a set of automatically generated rules in order to find a quantitative description for the data clusters [19]. In Fig. 4, the clustered SOM (a) and the descriptive rules for the corresponding clusters are shown (b). The map consists of 7 clusters, each shown using a different gray-level. In addition, the map units are labeled according to the cluster in which the map units belong to. On the right are shown the descriptive rules that indicate the kind of data samples in the different clusters. From the rules it can be seen that, for
E. Uplink results in microcellular scenario For the analysis of the uplink direction in the microcellular scenario, three variables have been selected. Selected 4
example, data cluster 6 represents data samples with an unacceptably high ulFER. It should be noted here that the automatically generated rules represent the same information as the component plane representation shown in Fig. 3 although in quite different form.
(a)
(b)
(c)
(d)
tion into behavioral cluster 3. In Fig. 5b, the geographical locations of mobile cells and their corresponding classification results are shown. The dominant behavioral clusters in terms of the number of classified base stations are 1, 2 and 6. Behavioral clusters 1 and 2 are described by rules for data clusters 3 and 5, and for behavioral cluster 6 the rules for data clusters 2, 3 and 5 of Fig. 4b apply. Typical of these behavioral clusters is the relatively low number of users, good quality and low uplink noise raise value. The explanation for the low number of users is the relatively small cell dominance area, which is typical of microcellular networks. When comparing the geographical area of cells in behavioral clusters 1, 2 and 6, high correlation can be found with uplink loading. (See Fig. 7 produced by the traditional analysis. The light areas in the figure indicate low loading.) The only cell in the bad performance area, that is, the behavioral cluster 3 described by rules for data cluster 6 and 7, is cell 44. Characteristics of this cell are high load and high number of users. The FER performance of this cell is degraded and thus, it can be concluded that the cell is operating at the edge of its capability. It is worth noting that an analysis utilizing conventional means did not identify cell 44 as problematic (see Sec. IV-B). Only the use of SOM and further analysis using expert knowledge proved this. In Fig. 6, the trajectories of mobile cells 8, 14 and 44 that can be used for mobile cell monitoring purposes are shown. Cell 8 has been chosen as an example to demonstrate the behavior of an ”average”cell, which is surrounded from all directions with interfering cells (see Fig. 5b). Compared to the situation with cell 44, the difference is that this cell is better isolated from the surrounding cells, and thus, the performance tends to be better. Cell 44 is surrounded by water areas and interfering signals can freely propagate. Whereas in the case of Cell 8, the street canyon structure and buildings isolate Cell 8 from the interfering sources. Cell 8 starts from behavioral cluster 2 with almost all measured data points in data cluster 5 indicating very low ulANR and ulFER. Then the operation point moves into behavioral cluster 6 due to an increase in nUsr and ulANR, because the number of data samples in data clusters 2 and 3 has increased. Behavioral cluster 8 represents another increase in the number of data samples with the number of users up to 6-10 and ulANR between 2.4-9.1 dB. Then the operation point visits behavioral cluster 1 with data samples mostly in data clusters 3 and 5, which in turn indicates operation with low number of users. In the end, the operation point on the histogram map returns back to behavioral cluster 6. The behavior in this cell as a function of time changes rapidly, but the number of users and loading are strongly correlated, and thus, the interference from other cells does not dominate the loading. The loading is generated by the traffic in cell 8 itself, and thus, no capacity is wasted. Cell 14 operates mostly in behavioral clusters 1 and 2 with almost all of the data samples in data clusters 3 and 5. As mentioned earlier, these data clusters represent data samples that have low nUsr and ulANR. The behavior of
Fig. 4. (a) K-means clustering of the map of data vectors, (b) the single variable rules for data clusters and (c-d) K-means clustering of the histogram map.
In Fig. 4, the clusters of the histogram map (c) and the hit-histogram prototypes stored in each map unit of the histogram map are shown (d). The histogram map consists of 8 behavioral clusters (behavior characterized by a histogram), each indicated by a different gray-level. The map units describing the cell behavior are labeled according to the behavioral cluster in which they belong. From the figure it can be seen that behavioral cluster 3 on the histogram map consists of feature vectors in which most of the single data samples are located in data cluster 6. This can be seen from the sixth bar of the hit-histograms in the corresponding behavioral cluster. Since data cluster 6 represented data samples with high ulFER (see Fig. 4b), the behavioral cluster 3 characterizes a cell behavior as undesirable. In Fig. 5a, a classification of mobile cells using the histogram map for uplink direction data is shown. The classification is based on histogram features which are computed from a time window. From the figure it can be seen that mobile cell 44 has quality problems due to its classifica5
cells.
Cell 8
Cell 14
Cell 44
Fig. 6. Trajectories of the cells.
In Sec. IV-B, the cells with interference problems were identified by traditional means. These cells are 3, 6, 7, 18, 24, 28, 42 and 43. When checking the position of these cells in Fig. 5a, one can see that they are dominantly located in behavioral clusters 2, 4 and 7. Behavior in clusters 4 and 7 is rather similar as it is characterized with relatively high noise rise, but moderate number of users. This finding supports that the cause is the interference from other cells. Behavioral cluster 2 is characterized by a low number of users, and thus, the other cells’ interference is again dominating in the load factor.
(a) 1
1 1
1
8
7
2
6 7
6
6
8
1
1
2
6 1
1 6
8
4
A. Analytical approach for the network performance data
6
5
1
2
IV. Conventional analysis of WCDMA cellular network
1
1
6
6 4
4
6
1
2
7
2
6 4 1
1
2
3
7
As presented in [20], the uplink load factor ηUL can be calculated as a sum of load factors of all N uplink connections in a cell. The effect of the multicell environment is taken into account by multiplying the single cell case with the term (1 + i), where i is other to own cell interference ratio. Thus, the loading for a single service, fixed mobile station (MS) speed, multicell case is
7
6
6
(b) Fig. 5. (a) Mobile cell clustering and (b) locations of classified cells.
ηUL = N cell 14 can be explained with the geographical position of this cell. It is at the edge of the analyzed area and thus, it has less interfering neighbors than the other cells. Low level of interaction between the cells and low number of users explain the rather static behavior of this cell. The behavioral clusters in question are characterized as ones having relatively low loading and good quality performance. Cell 44, located on an island off the coast of the city, first operates in behavioral clusters 7 and 8 with data samples in data clusters 1, 2, 4 and 7. This is the lower left corner of the map in Fig. 4b which represents data samples with different loads but low ulFER. However, the operation point moves into behavioral cluster 3 due to an increase in the number of samples in data cluster 6, that is, the cluster with the highest values for ulFER. Characteristic of this cell is that the loading ranges (both uplink and downlink) from moderate to high, but the number of users remains insignificant. The low number of users and high loading indicates interference problems due to interference from other
1 (1 + i) W 1 + ρR
(4)
in which R is the used bit rate, ρ is the signal-to-noise ratio requirement and W is the WCDMA chip rate. B. Traditional analysis results for microcellular case Using Eq.(4), two reference figures for the cell performance can be calculated based on the input data, namely the loading caused by a user and the number of users a cell can serve: The used uplink Eb /No value in this study was 3.5 dB. During the simulations loading was set to 0.95. Frequently quoted i value is 55%. The theoretical capacity values are presented in Table III. In addition to these analytical values the network was analyzed with a static radio network planning tool. These results from the radio network planning tool and especially the format used in visualization of the results is what the network management systems at their best can offer today: visualization support of individual KPIs and textual reports containing measurements and alarms. 6
TABLE III Theoretical capacity values.
Number of users, upper bound Loading/user, upper bound Number of users, i included Loading/user, i included
seen that the situation is not straightforward. Typically, admission control limits the number of users if the loading caused by active users is high. In this case, loading seems not to be the reason for blocking. Blocking is caused by downlink power outage, i.e. the maximum power allowed for one individual user. This explanation is based on the fact that microcellular scenarios are often downlink limited. In Table IV example results for the microcellular case are introduced. Only some example cells are chosen, the cells are the same as in the SOM trajectory analysis, i.e., cells 8, 14, 44. Full set of results can be found in Appendix in Table V.
26 0.03621 16 0.05613
BS14
BS1 BS13
BS2
BS15 BS3
BS35
BS4
BS16
BS24
BS34
BS39
BS32
BS23 BS17
BS10
BS40
BS33
BS26 BS11 BS5
TABLE IV Example results from traditional analysis.
BS25
BS12
Microcells
BS36
BS38
BS27
BS18
BS31
BS37 BS41
BS6
BS9
BS22
BS28
BS19
BS30
BS20
BS21
Cell Id BS Txp Traffic [W] Loading Other to own cell interference ratio, i Users UL Throughput UL [kbit/s]
BS44 BS46
BS43
BS8 BS7
BS42
BS45
BS29
(a)
BS15 BS25
BS12
BS35
BS4
BS16
BS24 BS26
BS34
BS17
BS10
BS36 BS27
BS18 BS6
BS39
BS32
BS23 BS5
BS40
BS33
BS11
BS31
BS38
BS37 BS41
BS9
BS22
BS28
BS19
BS30
BS20
BS21
BS42
BS44 BS46
BS43
BS8 BS7
BS45
BS29
22 18 22 1408 1152 1408
In Eq.(5), ThroughputN ORM is the normalized throughput. In the normalization, the maximum throughput was the maximum value in a cell. According to this classification the top 10% performing cells were cells 8, 9, 11, 25, 29 and 44.
Cell loading %
0.55
44 0.39 0.88 0.15
All of the selected microcells have relatively high loading, and significantly low i, compared to the often used value 55%. This indicates very good isolation of the cells. For all of these cells the loading per user is 0.04, which is very close to the upper bound value in Table III. In general, all the microcells have a well-controlled interference situation, only 8 cells out of 46 had i higher than 55%. These cells are: 3, 6, 7, 18, 24, 28, 42, and 43. For uplink performance PUL evaluation a simple function f combining the interference control and throughput aspects were generated, see Eq.(5). The weighting for each item in the cost function was the same. PUL = f (ThroughputN ORM , i, ηUL ) (5)
BS13
BS3
0.5
14 0.25 0.75 0.13
BS14
BS1 BS2
8 0.40 0.88 0.16
0.6
0.65
0.7
0.75
0.8
0.85
0.9
0.95
(b) Fig. 7. (a) Used network scenario and (b) Uplink loading. Site distance is couple of hundreds of meters. In (a) users with transmission power problems are denoted with ⋄ and ⋆ for uplink and downlink respectively. In (b) the lighter the gray-level the lower is the uplink loading.
V. Evaluation of the SOM based method A. Validity of the SOM results During analytical analysis (Sec. IV-B) for the microcell uplink case, the best performing cells were 8, 9, 11, 25, 29 and 44. When mapping these cells on the performance spectrum of Fig. 5a it is very interesting to note that cells 8, 9, 11 and 29 are all in the same behavioral cluster (i.e. 6). Cell 25 is in behavioral cluster 2, owing to the fact that as an edge cell (dominance area mostly on water) it has a lot less users than the other cells. Traditional means are not capable of finding the performance problems of cell 44. As noted earlier in Sec. III, this cell is performing at
In Fig. 7b, the uplink loading is visualized. The darker the gray-level the higher is the loading in a cell. In Fig. 7a, the location of the mobile stations that suffer from power outage are depicted. When uplink and downlink information is compared, it can be noted that the locations tend to correlate. Thus, uplink performance and downlink performance are well in balance. When the correlation between loading and the number of uplink users is tested it can be 7
the edge of its capabilities. Similar results can be found in the macrocellular and microcellular case when both the uplink and the downlink direction are analyzed [21]. As a conclusion, it can be said that traditional means support the conclusions of the cell classification performed by the SOM.
solutions that actually disturb the overall quality of the cell.
B. Convenience and usability of the SOM based analysis Current network performance monitoring and analysis tools are not capable of meeting the needs and requirements of service driven networks. The reason for this is the increasing number of measurements that one should process simultaneously. The introduction of each new service or service class will increase the amount of measurements the operator needs to collect from the network elements. Fig. 8 represents an example of a typical KPI output. Each measurement is presented separately and the end user is responsible for correlating the different measurements. Furthermore, there is also a physical limitation as to how many measurements can be visualized at once. The end user is responsible for combining the information from different domains.
Fig. 9. Measurement example from NMS with different time resolutions. For the data a minute, an hour and daily average filter has been used.
Deriving the QoE from solely objective measurements requires advanced intelligence and tools, like the SOM. The method proposed in this paper is highly visual and it can be used to combine different types of information like time and geography with the cell behavior profile. This is different from the traditional analysis where only a few KPIs are analyzed separately at a time. C. Scalability of the method As mentioned earlier, the network performance is currently handled by visualizing the KPIs separately on the screen. As a result, the number of graphs to be visualized increases w.r.t the number of services, the number of network elements and the number of KPIs used in the analysis. Thus, the optimization of the total performance over a large set of network elements, covering also all of the most important quality aspects, requires the analysis of a large number of separate graphs, and the combination of their results to form a single big picture of the current state of the network. However, the number of required visualizations does not increase so radically in the SOM-based approach presented in this paper. For example, the increase in the number of network elements increases only the number of cell trajectories (as in Fig. 6). If the number of services or the number of network measurements (variables) increases, the number of component plane visualizations as in Fig. 3 also increases. If only the most descriptive combination of the single variable rules for each data cluster is shown, the number of symbolic rules per data cluster increases with the number of used variables only if the new set of variables is able to separate the data clusters better than the old set of variables. Thus, the number of rules per cluster does not necessarily grow at the same rate as the number of variables used in the analysis. The SOM-based method is more scalable with respect to the changes in the problem setting than the currently used methods.
Fig. 8. An example of measurement visualization in network management system. For each measurement own window is needed.
Current methods rely strongly on averaging or cumulating values over longer period of time, most often one day. KPI values are analyzed as snapshots representing one period. This approach loses details such as the form of value distribution of a KPI inside the period. The approach can be enhanced by dividing the period into sub-periods and calculating averages or cumulative sums over those. This easily generates an amount of data so large that details vanish into them and the simultaneous analysis of KPI combinations or performance of cell group becomes even harder. Fig. 9 presents an example of NMS level KPI visualization. The same KPIs are presented with minute, hour and daily average filters. When a problematic KPI is detected, and it is further analyzed, it can happen that the actual behavior pattern of the cell and its development is not detected at all but the operator ends up optimizing or fixing only one aspect of the cell’s behavior. This can lead to sub-optimal 8
D. Computational complexity of the method The method consists of two independent algorithms (SOM, k-means) whose computational complexity is dependent on the amount of analyzed data. According to [10], the computational complexity of the SOM is proportional to N M d, where N is the number of data samples in the data set, M is the number of map units in the SOM and d is the number of variables (dimension) of the data set. Thus, the training of relatively small SOMs (less than thousand map units) is computationally efficient even for large data sets. In [13], the computational complexity of the k-means algorithm and its application to the clustering of the SOM codebook vectors is described. The computational complexity of the basic P k-means clustering algorithm is said to max be proportional to C k=2 N k, where N is the number of data samples and Cmax is the maximum number of clusters in which the data is to be clustered. However, when the kmeans clustering is applied to the codebook vectors of the SOM trained with the data set instead of the original data set, the computational PCmax cost of clustering is reported to reduce to N M + k=2 M k, thus making it computationally tractable solution for the analysis process described in this paper. As a reference, the computation time for a single network scenario was less than an hour, including the basic computation, visualization and report generation. VI. Usage of clustering in optimization The analysis method presented and used in this article consists of two different phases: the clustering of single data points consisting of several KPIs and clustering of sequences of these data points. The first phase is actually the traditional way to use the SOM. It is able to show what types of KPI value combinations there are in general. The basic clustering can be analyzed further in order to see how an average behavior of cells has developed during a longer period of time. In the case of behavioral clusters, the method takes into account not only the average behavior pattern based on all selected KPIs but also the recent variation in behavior pattern. By doing this, the method is able to detect the collapse of an excellent behavior much earlier than by analyzing only the average behavior. Cell clusters that have been found by the proposed methods can be used as a starting point for a more detailed analysis. This is especially suitable for trouble shooting type of tasks. For example, in Fig. 10, it can be assumed that the lower left corner of the SOM in the upper figure indicates cells that have a certain defect. These cells can automatically be selected as optimization targets. More detailed analysis can be provided on them and probably also actions to fix them can be suggested. The ability to combine the result provided by the SOM and the geographical locations of the cells is essential for an operator. The geographical location information can provide support to problem resolution. The cell clusters can also be used to simplify the task of parameter provisioning and optimizing. It is not feasible technically nor timewise to optimize the network cell by cell. Cell clustering can be used to increase the operational efficiency of an operator’s tasks. The cells can be
Fig. 10. Example of utilization of cluster information in optimization process. Selection for cells to be optimized/autotuned.
clustered based on their traffic profiles and density, propagation conditions, cell types, radio resource management functionality performance etc. When homogeneous meaningful groups have been found, all the cells in one group can share a set of configuration parameter values. This kind of grouping based on multiple criteria is more accurate and the operation of the network will benefit from this: the optimization process is greatly simplified, improved and made less error prone and more efficient. In behavioral clustering, the performance change over time can be shown on the SOM (see Fig. 6). The method actually combines two sources of information: change over time and type of behavior at each separated point of time. This type of analysis can be performed using data averaged over various time periods, ranging from tens of seconds to days. One could, for example, follow one cell movement in the SOM during peak traffic hours, assuming that networks are able to report cell performance frequently enough. The sequence clustering method can be applied to follow up the optimization actions made in the network. When the network element configuration is changed, the operator normally wishes to see the effect of the change to performance. The method can be used to show how the changed cells change their places on the SOM. This kind of infor9
mation combination will be essential in the optimization of WCDMA radio networks. The complexity of the radio networks is growing as well as the size of networks themselves. Operators will need means to rapidly analyze the changes in the network, given the high number of cells, several services with different QoS criteria and a large amount of collected performance data.
Acknowledgment The authors wish to thank Albert H¨oglund, Jukka Henriksson, Ari H¨am¨al¨ainen, Mikko T. Toivonen, Hannu Multim¨aki and other colleagues from Nokia and Sampsa Laine from Helsinki University of Technology for their valuable comments. References
VII. Conclusion
[1]
This paper proposes the use of the Self-Organizing Maps (SOM), a neural network method, in the analysis phase of high level telecommunication network optimization process. The results provided by the SOM were verified with the combination of traditional means and expert knowledge. It can be stated that the results show a good agreement. Being able to understand the SOM results with traditional means increases confidence in the novel analysis and its applicability in the area of cellular networks. During the course of this work it was noticed that the traditional analysis as such is not adequate enough to provide as enhanced demonstration of the network performance as the SOM provides. Current performance analysis methods include combinations of planning tool analysis data and real network measurements. The information format in a planning tool is a static snapshot of the network situation, and thus, the aspect of time is not present. The KPIs offered by the NMS are averaged values. Certain amounts of information of the past behavior of the KPI is also available. In the case of KPI information, no prior knowledge on the correlation of the different KPIs is available. The correlation analysis and the combination of these two information sources is done manually and it strongly relies on the expertise of the person performing the task. The strength of the SOM based analysis methods is in the fact that multiple measurements are used in the analysis at the same time. Furthermore, the output can be provided in a descriptive format to ease the operator decisions. The SOM based application in the analysis of cellular networks is not widely spread, and it is worth noting that the SOM based analysis has been successfully applied to GSM data to detect anomalous behavior of base stations, see [22] for more details. In the future, the operation of cellular networks will be strongly service driven. Compared to the current situation with provisioning of voice and simple best effort data services only, the change is enormous. Effective analysis of 2G networks’ voice service is currently challenging enough because of large volumes of data collected from network elements. The evolution towards 3G systems will further increase the amount of data. The operators’ task is to filter the relevant information to a level on which it can be easily handled. Furthermore, the data set must include all the essential parts needed to conclude the service quality. Operators can benefit from the introduced neural methods already today. The full gain and potential of the advanced analysis methods can be reached when multiple end-user services are provided, and the quality perceived by the customers need to be monitored and optimized.
[2] [3]
[4]
[5] [6]
[7] [8]
[9]
[10]
[11] [12] [13] [14] [15] [16]
[17] [18] [19]
[20] [21]
10
Telemanagement Forum, “Enhanced telecom operations map eTOM, The Business Process Framework for the Information and Communications Services Industry, Version 3.0,” Tech. Rep. GB921, June 2002. T. Kohonen, Self-Organizing Maps, Springer-Verlag, Berlin, 1995. K. Raivio, O. Simula, and J. Laiho, “Neural analysis of mobile radio access network,” in IEEE International Conference on Data Mining, San Jose, California, USA, November 29 - December 2 2001, pp. 457–464. P. Lehtim¨ aki, K. Raivio, and O. Simula, “Mobile radio access network monitoring using the self-organizing map,” in European Symposium on Artificial Neural Networks, Bruges, Belgium, April 24 - 26 2002, pp. 231–236. T. Kohonen, E. Oja, O. Simula, A. Visa, and J. Kangas, “Engineering applications of the self-organizing map,” Proceedings of the IEEE, vol. 84, no. 10, pp. 1358–1384, October 1996. E. Alhoniemi, J. Hollm´ en, O. Simula, and J. Vesanto, “Process monitoring and modeling using the self-organizing map,” Integrated Computer Aided Engineering, vol. 6, no. 1, pp. 3–14, 1999. O. Simula, P. Vasara, J. Vesanto, and R.-R. Helminen, Industrial Applications of Neural Networks, chapter The Self-Organizing Map in Industry Analysis, pp. 87–112, CRC Press, 1999. O. Simula, J. Ahola, E. Alhoniemi, J. Himberg, and J. Vesanto, “Self-organizing map in analysis of large-scale industrial systems,” in Proceedings of the Workshop on Self-Organizing Maps, Espoo, Finland, July 1 - 3 1999, pp. 375–387, (invited paper). T. Kohonen, J. Hynninen, J. Kangas, J. Laaksonen, and K. Torkkola, “SOM PAK: The self-organizing map program package,” Tech. Rep. A31, Helsinki University of Technology, Laboratory of Computer and Information Science, 1996, Available: http://www.cis.hut.fi/research/som pak/. J. Vesanto, J. Himberg, E. Alhoniemi, and J. Parhankangas, “SOM toolbox for Matlab 5,” Tech. Rep. A57, Helsinki University of Technology, Laboratory of Computer and Information Science, 2000, Available: http://www.cis.hut.fi/projects/somtoolbox/. B.S. Everitt, Cluster Analysis, Arnold, 1993. D.L. Davies and D.W. Bouldin, “A cluster separation measure,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 1, no. 2, pp. 224–227, April 1979. J. Vesanto and E. Alhoniemi, “Clustering of the self-organizing map,” IEEE Transactions on Neural Networks, vol. 11, no. 3, pp. 586–600, May 2000. K. Heiska and A. Kangas, “Microcell propagation model for network planning,” in Personal, Indoor and Mobile Radio Communications, 1996, vol. 1, pp. 148–152. J. Rajala, K. Sipil¨ a, and K. Heiska, “Predicting in-building coverage for microcells and small macrocells,” in IEEE Vehicular Technology Conference, 1999, vol. 1, pp. 180–184. S. H¨ am¨ al¨ ainen, H. Holma, and K. Sipil¨ a, “Advanced WCDMA radio network simulator,” in Personal, Indoor and Mobile Radio Communications, Osaka, Japan, September 12-15 1999, vol. 2, pp. 951–955. “Guidelines for evaluation of radio transmission technologies for IMT-2000,” 1997, Recommendation ITU-R M. 1225. “http://www.3gpp.org/,” . M. Siponen, J. Vesanto, O. Simula, and P. Vasara, “An approach to automated interpretation of SOM,” in Advances in Self-Organizing Maps, N. Allinson, H. Yin, L. Allinson, and J. Slack, Eds. 2001, pp. 89–94, Springer. J. Laiho, A. Wacker, and T. Novosad, Eds., Radio Network Planning and Optimisation for UMTS, John Wiley & Sons Ltd., 2001. J. Laiho, K. Raivio, P. Lehtim¨ aki, K. H¨ at¨ onen, and O. Simula, “Advanced analysis methods for 3G cellular networks,”
Tech. Rep. A65, Helsinki University of Technology, Laboratory of Computer and Information Science, 2002. [22] A.J. H¨ oglund, K. H¨ at¨ onen, and A.S. Sorvari, “A computer hostbased user anomaly detection system using the self-organizing map,” in Proceedings of the International Joint Conference on Neural Networks, 2000, vol. 5, pp. 411–416.
Appendix
11
TABLE V Results from traditional analysis Cell Id BS Txp Traffic [W] Loading Other to own cell interf. ratio, i Own cell loading ∆ loading caused by other cells [%] Users UL Loading/user Own cell interf./user Throughput UL [kbit/s] Cell Id BS Txp Traffic [W] Loading Other to own cell interf. ratio, i Own cell loading ∆ loading caused by other cells [%] Users UL Loading/user Own cell interf./user Throughput UL [kbit/s] Cell Id BS Txp Traffic [W] Loading Other to own cell interf. ratio, i Own cell loading ∆ loading caused by other cells [%] Users UL Loading/user Own cell interf./user Throughput UL [kbit/s]
1 0.39
2 0.24
3 0.46
4 0.56
5 0.28
6 0.15
7 0.15
8 0.40
9 0.41
10 0.23
11 0.28
12 0.46
13 0.32
14 0.25
15 0.30
16 0.24
0.84 0.32
0.72 0.51
0.89 0.88
0.89 0.20
0.70 0.28
0.65 0.65
0.65 0.59
0.88 0.16
0.89 0.07
0.89 0.29
0.89 0.12
0.89 0.53
0.89 0.28
0.75 0.13
0.89 0.40
0.73 0.32
0.63 24
0.47 34
0.47 47
0.74 17
0.55 22
0.39 40
0.41 37
0.76 14
0.83 7
0.69 23
0.79 11
0.58 35
0.69 22
0.66 12
0.64 29
0.55 24
19 12 16 22 15 11 12 22 25 21 23 17 21 18 19 15 0.044 0.060 0.055 0.040 0.047 0.059 0.054 0.040 0.035 0.042 0.039 0.053 0.042 0.042 0.047 0.049 0.033 0.040 0.029 0.033 0.036 0.035 0.034 0.035 0.033 0.033 0.034 0.034 0.033 0.037 0.033 0.037 1216
768
1024
1408
960
704
768
1408
1600
1344
1472
1088
1344
1152
1216
960
17 0.32
18 0.14
19 0.40
20 0.38
21 0.40
22 0.24
23 0.39
24 0.33
25 0.51
26 0.21
27 0.19
28 0.11
29 0.37
30 0.49
31 0.43
32 0.14
0.89 0.23
0.70 0.68
0.89 0.18
0.74 0.27
0.76 0.21
0.85 0.36
0.88 0.19
0.89 1.20
0.87 0.13
0.72 0.39
0.81 0.28
0.60 0.89
0.86 0.12
0.87 0.28
0.88 0.27
0.52 0.46
0.73 18
0.42 40
0.76 15
0.58 21
0.63 18
0.62 27
0.74 16
0.40 55
0.77 12
0.52 28
0.63 22
0.32 47
0.77 11
0.68 22
0.69 21
0.36 31
22 12 22 16 18 18 21 16 22 15 18 9 23 20 21 10 0.040 0.058 0.041 0.046 0.042 0.047 0.042 0.056 0.039 0.048 0.045 0.067 0.038 0.044 0.042 0.052 0.033 0.035 0.034 0.037 0.035 0.035 0.035 0.025 0.035 0.035 0.035 0.035 0.033 0.034 0.033 0.036 1408
768
1408
1024
1152
1152
1344
1024
1408
960
1152
576
1472
1280
33 0.38
34 0.22
35 0.23
36 0.28
37 0.34
38 0.36
39 0.22
40 0.17
41 0.35
42 0.12
43 0.16
44 0.39
45 0.29
46 0.30
0.72 0.31
0.80 0.26
0.79 0.20
0.89 0.28
0.79 0.53
0.87 0.37
0.68 0.31
0.64 0.51
0.89 0.23
0.46 0.85
0.75 0.63
0.88 0.15
0.77 0.30
0.85 0.21
0.56 23
0.63 20
0.66 17
0.69 22
0.51 35
0.63 27
0.52 24
0.42 34
0.73 19
0.25 46
0.46 39
0.76 13
0.59 23
0.70 17
14 18 19 20 15 18 12 12 22 6 13 22 16 21 0.052 0.044 0.042 0.044 0.052 0.048 0.056 0.053 0.041 0.076 0.057 0.040 0.048 0.040 0.040 0.035 0.035 0.035 0.034 0.035 0.043 0.035 0.033 0.041 0.035 0.035 0.037 0.033 896
1152
1216
1280
960
1152
768
12
768
1408
384
832
1408
1024
1344
1344
640
